Mclover

every AI is just math. I animate the equations behind the models.

[Mclover @0xmclover]

https://t.co/UUGnzFUEBI

[Mclover @0xmclover]

@NobelPrize that idea became a cubic kilometer of antarctic ice with 5,160 sensors hung on strings down to about 2.5 km

[Mclover @0xmclover]

@anirbanbandyo yo! would be great to connect with you

[Mclover @0xmclover]

@Math_files a team of 48 people tested this in 2023 with 350,757 flips and got the same side 50.8% of the time, close to diaconis's 51%

[Mclover @0xmclover]

Microsoft poured about $13 billion into OpenAI, Amazon put $8 billion into Anthropic, and Nvidia became the first $4 trillion company - all to push around arrows nobody can see. Every word you type into GPT-3 becomes a list of 12,288 numbers, one arrow in a space with 12,288 directions. Up there geometry breaks: a ball's volume peaks at 5 dimensions and then collapses toward zero, and in 1,000 dimensions 99.996% of it sits in the outer 1% - an orange that's all peel.

That glitch is the secret behind ChatGPT, Claude and Gemini. In high dimensions two random arrows are almost always at right angles, so every concept - French, sarcasm, a city, a chemical - gets its own private lane. Chris Olah's team at Anthropic calls it superposition, a model cramming far more ideas than it has dimensions. Tomas Mikolov's team at Google saw the first hint in 2013, when king minus man plus woman landed right next to queen.

5 minutes. Bookmark & watch today. The trillion-dollar geometry no human brain can picture

[Mclover @0xmclover]

https://t.co/UUGnzFUEBI

[Mclover @0xmclover]

@Math_files near a sharp jump the sum overshoots by about 9% no matter how many waves you add. that's the gibbs phenomenon, the ringing you see around edges in jpegs

[Mclover @0xmclover]

Nvidia became the first 4 trillion dollar company in history selling chips that mostly do one thing - multiply inputs by weights and add them up. That same operation made The New York Times in 1958, when the US Navy showed reporters a 2 million dollar IBM computer and called it the embryo of a machine that would walk, talk, reproduce itself and be conscious of its own existence. The man behind it was Frank Rosenblatt, and his whole idea was one line of math - step(w · x + b). Every neuron inside OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini still starts with that line.

Eleven years later MIT's Marvin Minsky, one grade behind Rosenblatt at the same high school, co-wrote a book proving a single line can never solve a 4 point puzzle called XOR. Funding moved elsewhere, Rosenblatt drowned on his 43rd birthday in 1971, and neural networks stayed out of favor for more than a decade until Rumelhart, Hinton and Williams used backpropagation to stack those lines into layers. Google later paid 44 million dollars for Geoffrey Hinton's 3 person startup. One of those 3 was Ilya Sutskever, who went on to co-found OpenAI, and in 2024 Hinton won a share of the Nobel Prize in Physics.

5 minutes. bookmark & watch today. the 1958 formula behind a 4 trillion dollar industry

[Mclover @0xmclover]

https://t.co/UUGnzFUEBI

[Mclover @0xmclover]

@mathelirium run it longer and you hit fractional revivals: at 1/q of the revival time the state splits into q copies. would love to see 2, 3 and 4 component cats side by side

[Mclover @0xmclover]

@PhilosophyOfPhy ludvig lorenz published retarded potentials in 1867, the same year riemann's paper finally came out. maxwell's dynamical theory had been out since 1865

[Mclover @0xmclover]

@mathelirium did you check drag against experiments? at re = 100 a sphere should land around cd ≈ 1.09, quick way to see if the mesh near the wall is fine enough

[Mclover @0xmclover]

Google DeepMind built an AI that broke a 53-year-old math record, and that record sits under the $700 billion Alphabet, Amazon, Microsoft and Meta plan to spend this year and under every chip that turned Nvidia into a $4 trillion company

Volker Strassen shocked mathematicians in 1969 by multiplying two grids of numbers with 7 steps instead of 8, and that one saved step cuts a 1024 by 1024 matrix from over 1 billion multiplications to about 282 million

AlphaTensor turned it into a game on top of AlphaZero, the AI that mastered chess and Go, found 47 multiplications where Strassen needed 49 in one setting, made matrix math 10 to 20% faster on Nvidia and Google chips, then AlphaEvolve sped up a Gemini kernel by 23% and cut Google's Gemini training time by 1%

5 minutes of math to watch before you pay OpenAI or Anthropic for your next million tokens, so bookmark it and watch today because the formula at the end tells you exactly how many operations every single AI answer burns

[vixhaℓ @TheVixhal]

https://t.co/UOX9HBErXe

[Mclover @0xmclover]

@wtgowers he's right that "understanding" keeps moving to wherever the machines haven't reached yet. chess went through the same thing with "creativity"

[Mclover @0xmclover]

@matharium george brown at rca came up with the turnstile in the 1930s for early tv broadcasting

[Mclover @0xmclover]

@Math_files only about 1 in 7 numbers is happy. 143,071 of the first million reach 1, the rest get stuck in that 4 cycle

[Mclover @0xmclover]

@sama showing confidence and sources by default would help. hard to worship something that openly says it's 60% sure

[Mclover @0xmclover]

@mathelirium that's a magnetic mirror, and earth has a giant one. electrons in the van allen belts bounce back and forth between the north and south poles the same way

[Mclover @0xmclover]

@sciam the render already has plants growing all over the dome, so the bacteria are way ahead of schedule

[Mclover @0xmclover]

@Math_files "the only way to learn mathematics is to do mathematics" reads a lot heavier after 120 applications and two no's

[Mclover @0xmclover]

@littmath going by the trend line it's ~10k to ~23k a month in 9 years, about 10% a year, and population grows under 1%. i'd bet most of it is cs/ml going arxiv-first and more fields posting preprints in the first place

[Aethir Ecosystem @AethirEco]

Aethir’s vision for a powerful, collaborative ecosystem is here. The Aethir’s power-packed #EcosystemExplorer is LIVE 🔒

Visit 👉 https://t.co/tgi6R5m3DR and venture on an exciting journey with us. Discover all our incredible partners, projects, and customers in one place.

1/3